image
Home Journals Progress in Chemistry
Progress in Chemistry

Abbreviation (ISO4): Prog Chem      Editor in chief: Jincai ZHAO

About  /  Aim & scope  /  Editorial board  /  Indexed  /  Contact  / 
Online first  /  Current issue  /  All issues  /  Top access  /  RSS

Progress in Chemistry >

2026 , Vol. 38 >Issue 2: 181 - 193

DOI: https://doi.org/10.7536/PC20250704

Review

Machine Learning-Assisted Nanomaterial Design and Preparation

  • Xiaoyang Wang 1 ,
  • Yifang Zhao 2 ,
  • Chenyi Liu 1 ,
  • Leyan Fan 3 ,
  • Dejun Xue , 2, * ,
  • Guolei Xiang , 1, *
Expand
  • 1 College of Chemistry， Beijing University of Chemical Technology， Beijing 100029， China
  • 2 Tongfang Knowledge Network Digital Technology Co.， Ltd.， Beijing 100192， China
  • 3 College of Materials Science and Engineering， Beijing University of Chemical Technology， Beijing 100029， China
* （Dejun Xue）；
（Guolei Xiang）

Received date: 2025-07-01

  Revised date: 2025-09-15

  Online published: 2025-09-30

Supported by

Advanced Materials-National Science and Technology Major Project(2025ZD0619900)

National Natural Science Foundation of China(22575013)

Fold

Abstract

Recent advances in machine learning （ML） have demonstrated remarkable potential in revolutionizing the design，property prediction，and synthesis optimization of nanomaterials，facilitating a paradigm shift from traditional empirical approaches to data-driven methodologies in nanoscience. This review examines the research frameworks and cutting-edge developments in ML-assisted nanomaterial design and fabrication，with a focus on representative material systems，including zero-dimensional quantum dots，one-dimensional nanotubes，two-dimensional materials，and metal-organic frameworks （MOFs）. Key technical aspects such as data acquisition and feature engineering，supervised and unsupervised modeling，generative algorithms，and automated experimental platforms are critically discussed. Furthermore，we highlight emerging challenges and future directions，emphasizing the need for standardized databases，physics-informed ML models，and closed-loop experimental systems to enable intelligent and efficient nanomaterial development. This work provides a comprehensive methodological reference for the integration of ML in next-generation nanomaterial research.

Contents

1 Introduction

2 Machine learning application framework

2.1 Acquisition and standardized preprocessing of high-quality data

2.2 Representation methods and feature engineering for material structures

2.3 Model construction and training

2.4 Validation and generalization assessment

2.5 Performance prediction and material screening

2.6 Inverse design and generative structural optimization

3 Representative research progress

3.1 Zero-dimensional nanomaterials

3.2 One-dimensional nanomaterials

3.3 Two-dimensional nanomaterials

3.4 Metal-organic frameworks

4 Conclusion and outlook

Key words： artificial intelligence; machine learning; nanomaterial synthesis; feature engineering; intelligent synthesis

Cite this article

Xiaoyang Wang , Yifang Zhao , Chenyi Liu , Leyan Fan , Dejun Xue , Guolei Xiang . Machine Learning-Assisted Nanomaterial Design and Preparation[J]. Progress in Chemistry, 2026 , 38(2) : 181 -193 . DOI: 10.7536/PC20250704

1 Introduction

As a class of structure-dominated functional materials, nanomaterials exhibit physical and chemical properties significantly different from bulk materials due to their characteristics such as small size effects, high specific surface area, quantum confinement effects, and high surface activity[1-2]. These unique properties make them widely explored and applied in frontier fields such as catalysis[3], energy conversion and storage[4-7], biomedical imaging and therapy[8-12], electronic devices[13], and environmental governance[14]. The nanostructuring of materials not only endows them with highly tunable structures and properties but also reshapes the research paradigm of traditional materials science, becoming a key driver for interdisciplinary innovation in chemistry, energy, biology, and electronics. With the continuous advancement of research, the functional development of nanomaterials increasingly relies on the precise control of their multi-scale structural parameters, including variables such as size, dimensionality, morphology, surface/interface configuration, chemical state, defect type, and electronic structure. The complex and high-dimensional space formed by these structural variables determines the properties and functions of nanomaterials; therefore, controlling the structural parameters of nanomaterials through the development of synthesis, preparation, modification, and processing strategies is the foundation for fully exploiting their functions and applications[15-18].
Traditional synthesis pathways for nanomaterials are divided into two approaches: "top-down" and "bottom-up." The former refines macroscopic materials into nanostructures through physical methods such as ball milling, etching, and exfoliation, while the latter constructs nanostructures based on chemical reactions via nucleation and growth mechanisms.[19]. Although such experience-driven trial-and-error methods promoted the diversified development of nanomaterial synthesis technologies in the early stages, they still exhibit significant deficiencies in addressing the complex coupling relationships of structural parameters and high-dimensional variable spaces. On one hand, reaction parameters such as solvent composition, temperature, pH, and ligand type are highly coupled during the synthesis process, making it extremely challenging to analyze the impact of a single variable on structural evolution.[20-21]; on the other hand, metastable intermediates exist in many nanosystems, and conventional characterization techniques struggle to track their dynamic evolution processes, limiting the in-depth understanding of nucleation-growth mechanisms.[22]. Furthermore, traditional manual synthesis methods face bottlenecks in efficiency and resource utilization, making it difficult to support systematic exploration of the experimental parameter space. Therefore, the field of nanosynthesis urgently needs to introduce new technological paradigms, integrating intelligent algorithms, automated synthesis, and high-throughput characterization into a "design-synthesis-characterization" system to achieve a leap from "empirical exploration" to "rational creation."[23-24].
The rapid development of Artificial Intelligence (AI) and Machine Learning (ML) has provided new opportunities to address the aforementioned challenges[25-30]. In the field of nanomaterials research, addressing issues such as the complex structure-property regulation mechanisms, low experimental efficiency, and long development cycles during the design and synthesis of nanomaterials, ML provides a powerful tool for the efficient development and performance optimization of new materials by constructing a closed-loop system of "data acquisition-model prediction-experimental verification"[31-32]. Currently, the application of AI in nanomaterial-related research covers multiple directions. In terms of performance prediction, data-driven models can efficiently predict key indicators such as the thermodynamic stability, electronic structure, band gap, and adsorption energy of nanoparticles, significantly reducing the computational cost of traditional simulations[33]; in terms of structure generation and screening, advanced generative models such as Generative Adversarial Networks (GAN) and Variational Autoencoders (VAE) support the inverse deduction of material structures based on target properties[34-35], greatly expanding the possibility space for material design; in terms of synthesis path planning, the combination of Natural Language Processing (NLP) technology with high-throughput experimental data enables the automatic extraction of reaction conditions from literature and the intelligent optimization of synthesis parameters, improving the efficiency of material preparation[36]. Furthermore, AI-driven high-throughput automated platforms combined with microfluidics technology can complete thousands of experiments in a short time; by dynamically adjusting reaction parameters through algorithms such as Reinforcement Learning (RL) and Bayesian Optimization (BO), they have been successfully applied to optimize the synthesis paths of quantum dots[37], MOFs[38-42], and catalysts[43-46]and other nanomaterials. Graph Neural Networks (GNN) and Convolutional Neural Networks (CNN) are applied to structure-property relationship modeling and spectral image recognition, respectively, achieving high-throughput classification and performance evaluation of nanoparticles. At the data level, AI tools are being widely used to build structured knowledge graphs and experimental databases, strongly promoting the integration and standardization of domain knowledge. These synthesis strategies integrating computational intelligence provide new pathways to break through the limitations of traditional methods, holding promise for in-depth analysis and rational design of the structure-property mapping relationships of nanomaterials. Focusing on core application scenarios such as performance prediction, synthesis optimization, inverse structure design, and characterization data analysis, this paper summarizes the technical routes, typical material systems, key algorithmic methods, and challenges of representative studies, providing a systematic theoretical basis and methodological reference for future AI-based intelligent nanomaterial design[47-50].

2 ML Application Framework

Traditional material design and synthesis strategies rely heavily on accumulated experimental experience or first-principles simulations, making it difficult to effectively reveal the high-dimensional complex relationships between material structure and performance. In contrast, ML leverages experimental and simulation data to establish nonlinear correlation models between material composition, structure, process parameters, and physicochemical properties through statistical methods, significantly improving the efficiency of new material development and advancing the processes of screening and process optimization (Figure 1). The application framework of ML in materials research typically includes the following six modules, which together constitute the foundational system for intelligent material design[51-53].
图1 AI在纳米材料优化工艺中的流程图

Fig.1 Flowchart of AI in the optimization process of nanomaterials

2.1 High-quality data acquisition and standardized preprocessing

High-quality data is the cornerstone of building reliable ML models. Available data sources in materials research mainly include high-throughput experimental results[54], first-principles[55-59](such as DFT), computational databases[60](such as Materials Project, OQMD), literature data[61-69](automatically extracted via NLP techniques), as well as image data[70-73](such as TEM and SEM images) and spectral data[74-78](XRD, Raman, XPS, etc.). However, these multi-source data generally suffer from heterogeneity, unstructured formats, and missing information. They require preprocessing steps such as cleaning, standardization (e.g., unit unification, redundancy removal, imputation of missing values), and consistency verification to ensure the accuracy and generalization capability of subsequent modeling. Common effective data preprocessing strategies include dimensionality reduction, such as Principal Component Analysis (PCA), feature standardization (e.g., Z-score normalization), outlier detection, and anomaly correction.

2.2 Representation Methods and Feature Engineering for Material Structures

Initial material structures, such as molecular formulas, crystal structures, dimensions, and defects, are transformed into digital features recognizable by algorithms (feature vectors) through specific strategies; this process is known as feature engineering or representation learning.[79-82]. For molecular systems, common representation methods include SMILES encoding, molecular fingerprints (e.g., ECFP), Coulomb matrices, and Bag-of-Bonds. For crystal structures, descriptors based on interatomic distances (e.g., SOAP), symmetry functions (Behler-Parrinello), or graph structure encodings are often employed. In recent years, Graph Neural Networks (GNNs) have been widely applied due to their inherent suitability for handling interatomic connectivity. GNNs can transform molecular or crystal structures into graphs composed of nodes (atoms) and edges (bonds or adjacency relationships). When combined with models such as CGCNN, MEGNet, and ALIGNN, they can effectively capture complex structural configuration features and local chemical environment information, significantly improving the accuracy of material characterization.

2.3 Model Construction and Training

The selection of ML models mainly depends on the task objective and learning paradigm; mainstream algorithmic models are divided into the following four categories.
Supervised Learning[83](Supervised learning, SL): Used for structure-property prediction (such as band gap, conductivity, adsorption energy, stability, etc.). Common algorithms include Random Forest (RF), Support Vector Machine (SVM), Gradient Boosting Trees (XGBoost), and Multilayer Perceptron (MLP).
Unsupervised learning[84](Unsupervised learning, UL): Suitable for material clustering, classification, image segmentation, and structural pattern discovery. Common algorithms include K-means, DBSCAN, PCA, etc.
Generative Models and Reinforcement Learning[85]: Applied to structural inverse design and material search, including Generative Adversarial Networks (GAN), Variational Autoencoders (VAE), Deep Reinforcement Learning (DRL), and Markov Chain Monte Carlo (MCMC) methods.
Bayesian Optimization and Active Learning[86](Active learning, AL): Combines experimental/simulation closed-loop systems for iterative optimization, widely used for guided decision-making in high-throughput experimental platforms.

2.4 Validation and Generalization Evaluation

To ensure model performance and reliability, it is necessary to introduce control mechanisms such as cross-validation (e.g., K-fold cross-validation), bias-variance assessment, and overfitting prevention during the model training process[87]. Evaluation metrics typically include the correlation coefficient (R2), Mean Squared Error (MSE), Mean Absolute Error (MAE), and Root Mean Squared Error (RMSE). For scenarios with scarce samples, special consideration must be given to the model's generalization ability and robustness; uncertainty quantification methods such as Monte Carlo and Bayesian Neural Networks can be combined to provide prediction confidence. Furthermore, model interpretability (e.g., SHAP, LIME) is gradually becoming a research hotspot, used to reveal the influence weights of input variables on output results, thereby enhancing the model's physical rationality and scientific credibility.

2.5 Performance Prediction and Material Screening

Trained ML models can efficiently predict the performance metrics of massive candidate material structures, thereby enabling high-throughput screening[88], with efficiency far exceeding traditional trial-and-error methods. For example, in catalyst screening, ML models can predict reaction activity based on adsorption energy; in optoelectronic materials, models can evaluate the bandgap and carrier mobility of different materials. This process is typically combined with high-throughput material libraries (such as AFLOWlib and Materials Cloud) to enhance the efficiency of discovering new materials.

2.6 Inverse Design and Generative Structural Optimization

The core goal of inverse design is to inversely infer the optimal material structure that meets specific functional requirements based on target performance criteria. Generative models (such as GANs, VAEs, and Diffusion Models) can sample and generate novel structures from latent space, combine with performance predictors to filter optimal candidates, and be applied to scenarios such as SMILES generation for new molecules, MOF lattice design, and nanoparticle configuration prediction.[89-92]. For example, researchers have utilized DRL to automatically generate metal cluster structures with high catalytic activity and predicted their stability and synthetic feasibility through AI models, thereby guiding experiments.

3 Representative Research Progress

Over the past decade, machine learning (ML) has demonstrated immense application potential and achieved significant research progress in the field of nanomaterials, particularly in structural design, property prediction, and synthesis process control. Focusing on nanomaterial systems of different dimensions, researchers have extensively explored ML applications in representative materials such as zero-dimensional (e.g., quantum dots, metal nanoclusters), one-dimensional (e.g., carbon nanotubes, nanoribbons), two-dimensional (e.g., graphene, TMDCs), and metal-organic frameworks (MOFs), constructing various technical pathways that integrate property modeling, synthesis regulation, and inverse design. Following the structural systems of zero-dimensional, one-dimensional, and two-dimensional nanomaterials, and incorporating MOF materials, this article discusses representative research achievements in these related materials and outlines the development trends and major advances in this interdisciplinary research field.

3.1 Zero-dimensional nanomaterials

Zero-dimensional nanomaterials refer to particle systems in which all dimensions are confined to the nanoscale in three-dimensional space. Typical representatives include semiconductor quantum dots, metal nanoclusters, carbon quantum dots, and fullerenes. These materials often possess unique luminescent properties, size-dependent electronic structures, high surface activity, and biocompatibility, and are widely applied in fields such as optoelectronic devices, bioimaging, catalysis, sensing, and drug delivery.

3.1.1 Quantum dot materials

Quantum dots (QDs) are a class of zero-dimensional semiconductor nanomaterials with sizes approaching the sub-nanometer scale and exhibiting significant quantum confinement effects; their electronic energy levels are discrete, demonstrating size-tunable optical and electronic properties.[93]However, the synthesis of quantum dot materials, due to their performance being strongly dependent on particle size distribution, morphological uniformity, and surface chemical environment, often requires precise control over nucleation and growth processes within multi-parameter coupled reaction systems to achieve consistent control over structure and properties, imposing extremely high demands on the reproducibility of synthesis conditions and process stability.
To address the aforementioned issues, Shen et al.[94]proposed a methodological framework based on ML and real-time feedback control, successfully achieving precise regulation of InAs/GaAs quantum dot density and providing an intelligent solution for the controllable growth of zero-dimensional nanomaterials. This study, by combining molecular beam epitaxy (MBE) technology with reflection high-energy electron diffraction (RHEED) video data (Figure 2a), constructed a broad range spanning from zero density to high density (1.4×1011 cm-2). In terms of model design, the study innovatively employed a three-dimensional residual neural network (3D ResNet-50) (Figure 2b) to process RHEED video sequences, extracting spatiotemporal features of surface morphology changes, effectively breaking through the limitations of traditional two-dimensional image analysis. By constructing a dual-model system comprising a "QDs model" and a "density model," the study achieved high-precision discrimination of quantum dot formation states (accuracy of 94.4%) and their density levels (accuracy of 95.1%), and utilized model outputs for real-time temperature regulation, dynamically adjusting the quantum dot density from 1.5×1010 cm-2 to 3.8×108 cm-2 or 1.4×1011 cm-2.
图2 基于深度学习的材料图像分类流程示意图：（a） 样本图像的处理方法。（b） 模型的简化架构图。预处理后的数据由一个3D ResNet 50模型处理，该模型包含残差结构和全连接结构，并输出分类结果[94]

Fig.2 Workflow for deep learning-based material image classification. （a） Processing method of sample images. （b） Simplified architecture diagram of the model. The preprocessed data is processed by a 3D ResNet 50 model，which contains residual structures and fully connected structures，and outputs classification results[94]

This work deeply couples deep learning with the MBE process, constructing a closed-loop feedback mechanism based on real-time video data. It effectively replaces traditional trial-and-error methods, significantly enhancing the efficiency, stability, and reproducibility of nanostructure synthesis. In the future, combining RL to optimize multi-parameter synergistic control can drive the leapfrog development of zero-dimensional nanomaterials from targeted design to intelligent manufacturing.
Carbon quantum dots (CDs) are a class of zero-dimensional nanomaterials primarily composed of carbon, possessing excellent fluorescence properties, biocompatibility, and environmental friendliness, and are widely used in fields such as bioimaging, sensing, and anti-counterfeiting technologies.[95]However, their synthesis processes are mostly based on liquid-phase methods involving the synergistic regulation of multiple parameters. Furthermore, the luminescence mechanism of the products is closely related to structural heterogeneity, making it extremely difficult to achieve high quantum yield, uniform morphology, and controllable fluorescence characteristics.
To address this challenge, Han et al.[96]proposed an ML-driven optimization strategy (Figure 3a), achieving efficient regulation of the fluorescence quantum yield of carbon dots and demonstrating the application prospects of intelligent synthesis in the field of zero-dimensional nanomaterials. The researchers prepared 391 groups of CD samples via a hydrothermal method, establishing an experimental database covering key variables such as reaction temperature (100–220 °C), precursor mass (0.02–0.1 g), and ethylenediamine (EDA) volume (0–200 μL). They utilized an XGBoost regression model to investigate the complex nonlinear relationship between parameters and QY (Figure 3b), revealing that EDA dosage and precursor mass are the most influential factors (Figure 3c), with their combined contribution to the yield exceeding 60%. Through the optimal synthesis conditions predicted by the model (Figure 3d), green fluorescent CDs with a QY of 39.3% were successfully prepared, featuring a size of 2 nm and a single-layer graphene structure.
图3 用于指导 CDs 合成的ML应用。 （a） 基于ML和水热实验的高量子产率 CDs 指导合成的设计框架。 （b） 水热生长 CDs 的选定特征之间的皮尔逊相关系数矩阵的热图。 （c） 从学习完整数据集的 XGBoost-R 中检索到的特征重要性。 （d） 训练有素的模型的预测，由两个最重要的特征形成的矩阵表示[96]

Fig. 3 Machine learning applications for guiding the synthesis of CDs. （a） Design framework for guiding the synthesis of high quantum yield CDs based on machine learning and hydrothermal experiments. （b） Heatmap of the Pearson correlation coefficient matrix between the selected features for hydrothermal growth of CDs. （c） Feature importance retrieved from XGBoost-R trained on the complete dataset. （d） Predictions of the well-trained model，represented by a matrix formed by the two most important features[96]

Further performance tests indicate that the CDs exhibit excellent selectivity for Fe³⁺ ions, with a favorable linear response range (0–150 μmol/L) and a low detection limit (0.039 μmol/L). The innovation of this work lies in combining machine learning with traditional hydrothermal methods; through feature importance analysis and parameter space mapping, it overcomes the bottlenecks of conventional trial-and-error approaches in complex multidimensional optimization, offering a data-driven new perspective for the targeted design of zero-dimensional nanomaterials. In the future, integrating chemical composition features (such as doping types) with synthesis parameters could enable the construction of more comprehensive predictive models, driving a paradigm shift in carbon-based nanomaterials from empirical exploration to intelligent synthesis.

3.1.2 Metal nanoclusters

Coordination-type metal clusters are nanoscale aggregates with stable configurations formed through the self-assembly of elements such as gold, silver, and copper with organic ligands like thiols via coordination bonds. They often possess tunable luminescent properties and are extensively studied in fields including biological labeling, catalysis, sensing analysis, and light-emitting devices.[97]. However, their synthesis involves the combined influence of multiple factors such as metal core size, oxidation state, ligand type, and coordination mode, often facing challenges including low cluster structural stability, difficulty in controlling the nucleation-growth process, and complex structural characterization, which to some extent limits their scalable preparation and precise regulation of structure-performance relationships.
Panapitiya et al.[98]proposed an ML model based on the Random Forest (RF) algorithm to efficiently predict the adsorption energy of carbon monoxide (CO) on thiolate-coordinated gold-silver alloy nanoclusters (Au(25-x)Agx(SR)18) (Figure 4a), providing a low-cost, scalable computational solution for optimizing the catalytic performance of zero-dimensional nanomaterials. The study constructed a dataset containing over 2,000 samples based on DFT calculations, covering the alloying range from single silver atom doping (x = 1) to full silver substitution (x = 25). In feature engineering, to accurately describe the impact of cluster structure on adsorption performance (Figure 4b), the researchers systematically extracted four categories of structural features, including interatomic distances, bond counts, graph structure representations, and atomic enclosure volumes. Through feature selection, several key descriptors were identified (Figure 4c), such as the spatial distribution of silver atoms and the number of H-C-H fragments; these parameters played a crucial role in revealing the nonlinear relationship between adsorption energy and the local chemical environment.
图4 CO在Au25团簇表面的吸附研究。（a） CO/Au25系统。用虚线红色曲线表示两个最近邻层的假定边界。AS代表CO吸附位点。表面的吸附Au/Ag原子位点用蓝色表示。（b） 计算的CO吸附能的变化。（c） 随机森林选出的最重要的特征以及对应的皮尔逊相关系数和互信息值[98]

Fig.4 CO adsorption on the surface of Au25 clusters. CO/Au25 system. The assumed boundary between the two nearest neighbor layers is indicated by the dashed red curve. AS represents the CO adsorption site. The surface adsorption sites of Au/Ag atoms are shown in blue. （b） Variation of the calculated CO adsorption energy. （c） The most important features selected by the random forest，along with the corresponding Pearson correlation coefficients and mutual information values[98]

By introducing second- and third-order feature transformations (such as squared and cubed terms) and employing a two-stage feature selection mechanism, the model demonstrates good predictive performance across several typical cluster systems, including Au25, Au36, and Au133nanocluster systems, achievingR2 = 0.78, 0.65, and 0.75 prediction accuracy, significantly outperforming traditional empirical models based on coordination numbers. This method does not rely on structural relaxation; it achieves efficient and accurate adsorption energy predictions using only static features from unoptimized geometric structures, thereby significantly reducing computational costs. Future research could further integrate structural relaxation dynamics, multi-component adsorption coupling effects, and other factors, while introducing algorithms with stronger structural perception capabilities such as GNNs, to achieve more comprehensive predictions and precise design of nanocatalysts under real reaction conditions.
Mekki-Berrada et al.[99]proposed a two-stage ML framework based on the synergistic mechanism of BO and Deep Neural Networks (DNN), which was successfully applied to parameter optimization and precise prediction of optical properties in the synthesis process of silver nanoparticles (Ag-NPs). At the data processing level, the authors designed a composite loss function based on cosine similarity and spectral intensity to achieve simultaneous optimization of spectral shape and intensity (the optimization process and results are shown in Figure 5b), improving the stability of the fitting results. In terms of feature engineering, five reaction parameters, including flow rate ratio and total flow rate, were selected as inputs to output full-band spectral curves, avoiding limitation to a single indicator and reducing information loss. Regarding algorithm implementation, BO efficiently explores the parameter space during the early data-sparse stage, while DNN takes over the nonlinear function modeling task as data gradually accumulates, balancing global search capability with local fitting accuracy (Figure 5a). Furthermore, through SHAP (Shapley Additive Explanations) analysis, the contribution of key process parameters to optical properties was revealed, enhancing the model's interpretability. Ultimately, this framework successfully synthesized silver nanoprisms with an edge length of approximately 65 nm within only 120 experimental samples, and their structural precision was verified via TEM images (Figure 5c). This study provides an efficient and generalizable ML-assisted strategy for the inverse design of optical nanomaterials, demonstrating the practicality and advancement of data-driven methods in synthetic chemistry.
图5 机器学习在优化银纳米棱镜合成中的应用。 （a） 损失随时间的演变，由随机搜索（绿色）、贝叶斯优化（蓝色）和深度神经网络（橙色）建议的条件：每个点代表一个液滴。BO的吸收光谱（b） 二维映射显示了使用原始实验数据获得的最小损失。（c） 在银纳米棱柱合成中的知识提取[99]

Fig.5 The application of machine learning in optimizing the synthesis of silver nanoprisms. （a） Evolution of loss over time，suggested by random search （RS，green），Bayesian optimization （BO，blue），and deep neural networks （DNN，orange）：Each point represents a droplet. Absorption spectra for BO. （b） Two-dimensional mapping showing the minimum loss obtained using the original experimental data. （c） Knowledge extraction in the synthesis of silver nanoprism[99]

3.2 One-dimensional nanomaterials

Carbon nanotubes (CNTs) are one-dimensional hollow tubular structures formed by rolling graphene sheets. Due to their exceptional mechanical, electrical, and thermal properties, they hold broad application prospects in nanoelectronic devices, composite materials, and energy technologies.[100]However, the properties of CNTs strongly depend on their chiral structure, the formation of which is synergistically regulated by multiple factors such as catalyst morphology, reaction atmosphere, and growth kinetics. Therefore, achieving precise control over chirality remains a key challenge in the field of one-dimensional carbon material synthesis.
To address this issue, Sun et al.[101]developed a molecular dynamics simulation method based on a machine learning force field (MLFF), revealing the kinetic mechanism underlying the origin of chirality selection during single-walled carbon nanotube growth and providing new insights for directed synthesis. Focusing on the cobalt-carbon interface system, the study constructed and optimized the MLFF through an active learning strategy, achieving efficient cross-scale simulation of the catalyst–carbon precursor interface reaction process while maintaining first-principles accuracy. In terms of modeling strategy, the authors introduced a reaction network graph representation method based on a "0-1" edge pattern. Combined with microkinetic modeling, this approach elucidated the intrinsic correlation between chirality distribution and defect dynamics (Fig. 6a, b). Simulation results show that carbon nanotubes with (6,5) chirality exhibit a significant advantage during growth, highly consistent with the most common chiral products observed in experiments, thereby confirming the validity of the model. To further explain the chirality evolution phenomenon, the research team proposed two diameter control mechanisms (Type I and Type II) and systematically analyzed the impact of defect types and their evolutionary behavior on chirality stability. The results indicate that tricoordinate defects (l = 3) in near-armchair structures have faster repair rates, making them more prone to stable growth (Fig. 6c, d); whereas near-zigzag structures are difficult to maintain due to the poor stability of tetracoordinate defects (l = 4) (Fig. 6e).
图6 纳米管-催化剂界面小周长的螺旋度分布和观察到的直径控制机制。（a） “五层”分布；（b） “零层”分布；（c） 具有（7，2）螺旋度的I型机制。 （d） 具有（7，5）螺旋度的I型机制。壁中的额外五边形已成功愈合。 （e） 从（可能的）（8，4）到（8，3）的II型机制转变。包裹的五边形未愈合，形成一个七边形来补偿它。粉色（灰色）球体是钴（碳）原子，在图（c~e）中，五边形（七边形）分别用蓝色（红色）表示[101]

Fig.6 Distribution of chirality and observed diameter control mechanisms at the nanotube-catalyst interface for small perimeters. （a） “Five-layer” distribution. （b） “Zero-layer” distribution. （c） Type I mechanism with （7，2） chirality. （d） Type I mechanism with （7，5） chirality. The extra pentagon in the wall has successfully healed. （e） Type II mechanism transition from （possibly） （8，4） to （8，3）. The wrapped pentagon does not heal and forms a heptagon to compensate for it. Pink （gray） spheres are cobalt （carbon） atoms，and in figures （c~e），pentagons （heptagons） are represented in blue （red）[101]

This work reveals the kinetic origin of chirality selection in SWCNTs at the atomic scale, emphasizing the importance of defect management in chirality-directed growth. The study fully demonstrates the advantages of ML force fields in simulating the growth mechanisms of one-dimensional nanomaterials, providing theoretical tools and a simulation framework for the precise design of carbon-based nanomaterials.
Xu et al.[102]reported a machine learning-assisted chemical vapor deposition (CVD) synthesis method (the overall design framework of which is shown in Figure 7), successfully achieving the controlled preparation and geometric morphology regulation of one-dimensional (1D) nanomaterials—few-layer WTe2 nanoribbons (NRs), demonstrating the significant role of ML technology in the visual design and intelligent optimization of low-dimensional materials. This study constructed a supervised learning model (XGBoost) combined with 255 sets of experimental data to achieve intelligent optimization of key synthesis parameters (such as H2 flow rate, Te/W source ratio, etc.), thereby significantly improving the synthesis efficiency and morphological consistency of one-dimensional nanostructures.
图7 基于ML和实验实现几何控制的 WTe2化学气相沉积合成的设计框架[102]

Fig.7 Design framework for the synthesis of WTe2 chemical vapor deposition based on machine learning and experimental realization of geometric control [102]

Studies show that H2 flow plays a dominant role in the formation of WTe2 nanoribbons, while the Te/W source ratio (RTe/W) regulates geometric features such as the aspect ratio of the nanoribbons. By adjusting RTe/W (1~20), the nanoribbon width (0.7~1.9 μm) and length (15.7~49.2 μm) can be precisely controlled. Combining theoretical calculations with experimental observations, the study further reveals the formation mechanism of WTe2 nanoribbons: H2 atmosphere not only promotes the effective transport of Te but also induces the transformation of two-dimensional WTe2 crystals into one-dimensional structures through an edge-selective etching mechanism. This growth pathway from 2D materials to 1D nanoribbons achieves precise guidance of structural dimensionality via atmosphere control and source ratio adjustment, providing important insights for understanding the formation mechanisms of 1D nanomaterials. This work not only offers a new approach for preparing 1D nanostructures of the topological insulator WTe2, but also demonstrates the great potential of ML in the synthesis of low-dimensional nanomaterials, including parameter optimization, mechanism elucidation, and universality across material systems (e.g., extending to MoTe2, etc.).

3.3 Two-dimensional nanomaterials

Two-dimensional nanomaterials (2D materials) refer to nanosheet layered structures with single-layer or few-layer atomic thickness in the longitudinal dimension, capable of exhibiting unique electronic, optical, mechanical, and catalytic properties[103-105]. Typical two-dimensional nanomaterials include graphene, transition metal dichalcogenides (such as MoS2, WS2), hexagonal boron nitride (h-BN), and MXene. Benefiting from their atomic-level thickness and high surface exposure, these materials generally possess high specific surface area, tunable interlayer coupling effects, and excellent carrier mobility, demonstrating broad application prospects in flexible electronic devices, energy storage, electrocatalysis, biomedicine, and other fields. However, the synthesis and preparation of 2D materials face multiple challenges: whether through physical exfoliation or chemical vapor deposition methods, achieving large-area and uniform preparation remains difficult; in chemical synthesis pathways, precise control over layer number, defect density, edge configuration, and phase purity still poses difficulties, constraining systematic research on structure-property relationships and reproducibility in device integration.
Dahl et al.[106]proposed a methodological framework combining in situ spectral kinetics with Scientific Machine Learning (Scientific ML), systematically revealing the nucleation and growth mechanisms of two-dimensional perovskite nanoplatelets (OLA2PbBr4), providing theoretical and methodological support for the controlled synthesis of complex nanosystems. Through high-throughput experimental design, this study employed a stopped-flow experimental scheme (Figure 8a) to collect over 500,000 in situ spectral data points covering more than a hundred reaction conditions (typical experimental spectra as shown in Figure 8b), constructing a high-quality, standardized experimental database. In terms of theoretical modeling, the team established a mapping relationship between exciton energy and nanoplatelet size based on quantum mechanical calculations, effectively addressing the limitations of traditional kinetic methods in modeling within high-dimensional parameter spaces. Regarding algorithm design, the study adopted a differential evolution optimization algorithm for global optimization of kinetic parameters and enhanced the model's generalization performance through extrapolation cross-validation, successfully predicting two nanoplatelet formation pathways and their size evolution trajectories; the model fitting results showed good agreement with experimental data (Figure 8c), with their size evolution trajectories as shown in Figure 8e. The innovation of this work in multi-scale modeling lies in correlating macroscopic spectral features with microscopic structural changes, achieving coupled modeling between macroscopic and microscopic scales through a size-dependent optical model (Figure 8d), thereby enhancing the physical interpretability of the model. This framework can further incorporate algorithms such as GNN or RL to expand its support capabilities for reaction pathway search and real-time process control, promoting the construction of a closed-loop material design from data-driven approaches to intelligent decision-making.
图8 组合模型概述。（a） 停流方案，在透明窗口中混合苯甲酰溴和铅离子与油胺，观察纳米片的形成。 （b） 从停流测量中获得的吸收光谱的时间序列示例。 （c） 从组合模型和拟合数据 （b） 中获得的吸收光谱的时间序列示例。 （d） 作为粒径函数的消光系数的光学模型（蓝色，小到红色，大）。 （e） 不同尺寸纳米片浓度随时间变化的动力学模型[106]

Fig. 8 Overview of the combined model：（a） Stop-flow scheme，mixing benzoyl bromide and lead ions with oleylamine in a transparent window to observe the formation of nanosheets. （b） Example of a time series of absorption spectra obtained from stop-flow measurements. （c） Example of a time series of absorption spectra obtained from the combined model and fitted data （b）. （d） Optical model of the extinction coefficient as a function of particle size （blue for small to red for large）. （e） Kinetic model of the concentration of nanosheets of different sizes as a function of time[106]

3.4 MOF

Metal-organic frameworks (MOFs) are porous crystalline materials formed by the self-assembly of metal ions or metal clusters with organic ligands through coordination bonds. They feature high specific surface area, designable structures, and diversity, making them an important type of functional nanomaterial.[107-109]. MOF materials demonstrate broad application potential in fields such as gas storage and separation, catalysis, electrochemical energy storage, and biomedical applications.
However, the synthesis and preparation of MOF materials still face numerous challenges. The crystal nucleation and growth processes are highly sensitive to temperature, pH, and solvent systems, making it difficult to precisely control product particle size, morphology, and crystallinity. Coordination competition between different organic ligands and metal nodes can easily induce the formation of multiple isomers or topological structures, increasing the difficulty of material screening and reproducible preparation. Furthermore, most MOFs exhibit poor stability in aqueous or acidic environments, limiting their practical applications. Therefore, developing effective synthetic regulation strategies to achieve precise control over the structure, scale, and composition of MOFs, thereby constructing MOF nanomaterials with high stability and tunable functionality, has become an important research direction in this field.
Goeminne and Van Speybroeck developed a new method combining machine learning potentials (MLP) with transition matrix Monte Carlo (TMMC) simulations (a comparison with traditional methods is shown inFigure 9), achieving high-precision prediction of water adsorption behavior in MOF materials and providing a new pathway for regulating the adsorption properties of porous nanomaterials[110]. This study efficiently trained the MLP model through an active learning strategy, requiring only thousands of DFT calculations to cover the fully flexible configuration space of materials such as MOF-303, MOF-333, and MOF-LA2-1, significantly reducing computational costs.
图9 吸附建模方法的对比。传统的力场-大正则蒙特卡罗方法与本研究采用的机器学习势-转移矩阵蒙特卡罗方法[110]

Fig.9 Comparison of adsorption modeling approaches. the traditional force field-grand canonical Monte Carlo method versus the machine learning potential-transition-matrix Monte Carlo method employed in this study[110]

The constructed MLP model accurately captures hydrogen bond interactions between water molecules and the MOF framework through atomic environment descriptors, achieving a prediction error of less than 0.9 meV/atom, which meets the accuracy requirements close to quantum chemical calculations. By comparing the performance of five DFT functionals, the study confirms that the rPBE-D3(BJ) method provides the description of water-MOF and water-water interactions closest to CCSD(T) reference values, laying a theoretical foundation for model reliability. During the training data generation process, the authors sampled local and global framework deformations during the adsorption process via multi-temperature molecular dynamics simulations (100–500 K), ensuring the comprehensiveness and representativeness of the data. Regarding simulation strategies, TMMC directly samples flexible frameworks using molecular dynamics trajectories in the NPT ensemble, breaking through the limitations of traditional rigid framework assumptions and quantitatively reproducing the S-shaped features of the MOF-303 experimental isotherm for the first time. Further three-dimensional density distribution analysis reveals that the strong adsorption sites formed by pyrazole groups in MOF-303 serve as nucleation centers for water molecules at low pressures, whereas the introduction of hydrophobic FDC linkers in MOF-333 suppresses this effect, thereby optimizing the step characteristics of the isotherm. This work not only resolves simulation challenges arising from the coupling of framework flexibility and adsorption behavior but also establishes a generalizable computational framework for the inverse design and high-throughput screening of MOF-based nanomaterials, highlighting the critical role of machine learning in the precise prediction of adsorption properties in nanoporous materials.
He et al.[111]reported an ML-driven system based on the synergy between the XGBoost algorithm and a robotic synthesis platform, successfully achieving efficient discovery and synthesis optimization of polyoxometalate-based metal-organic frameworks (POMOFs), providing a new paradigm for the intelligent exploration of complex multi-parameter material systems. This study established a closed-loop experiment-modeling feedback process (its synthetic reaction space is shown in Figure 10a), utilizing real-time experimental results to drive model iteration, significantly improving the efficiency of material discovery and the reproducibility of experimental operations. At the data processing level, the authors employed an XGBoost model extended for multi-class classification, combined with an uncertainty-guided sampling strategy (Figure 10d), dynamically updating model parameters across different experimental rounds, ultimately achieving a crystal formation prediction accuracy with an F1 score exceeding 0.8, demonstrating excellent classification performance and robustness. In terms of feature engineering, six-dimensional parameters including ligand type, concentration ratio, reaction sequence, and temperature were selected as inputs to output crystal formation probability, and the influence of multi-dimensional parameters was visualized through chemical space mapping. At the algorithmic level, the XGBoost model, through iterative optimization and data augmentation, balanced global exploration capability with local prediction accuracy, while utilizing χDL (Chemical Description Language) to standardize synthesis processes, ensuring high experimental reproducibility. Electrochemical performance tests further indicated that the zinc ion ratio, ligand type, and topological structure have significant regulatory effects on the electron transfer capability of POMOFs. Ultimately, this study successfully discovered nine novel POMOFs, including one mixed-ligand structure, and verified their structural characteristics via single-crystal X-ray diffraction (Figure 10c). This study not only demonstrates the powerful modeling capability of ML in exploring high-dimensional synthesis spaces but also highlights the critical role of automated platforms and standardized experimental languages in promoting the reproducibility and scalable exploration of material synthesis, providing a highly integrated and generalizable research framework for the intelligent design and performance optimization of MOF materials.
图10 多金属氧酸盐金属有机框架（POMOFs）的机器人发现示意图：（a） 一锅法合成POMOFs的反应空间。（b） 用于进行反应的机器人平台。（c） 反应结果的确认和记录。（d） 用于优化ML模型的不确定性反馈[111]

Fig.10 Schematic of robotic discovery of polyoxometalate metal-organic frameworks （POMOFs）：（a） reaction space for one-pot synthesis of POMOFs，（b） robotic platform for conducting reactions，（c） confirmation and recording of reaction outcomes，and （d） uncertainty feedback for optimizing the machine learning model [111]

4 Conclusion and Outlook

With the rapid development of data science and AI technologies, ML techniques have demonstrated immense application potential in nanomaterial design, synthesis, and performance prediction, gradually becoming a core driver propelling the transformation of research paradigms in materials science. Currently, methods such as supervised learning-based performance prediction models, generative model-assisted inverse structural design, GNNs for processing atomic graph structures, and BO-guided experimental pathway search have been progressively integrated into nanomaterial research workflows with initial success. Meanwhile, large language models represented by ChatGPT are also being utilized for literature mining and data extraction, supporting the automated construction of material knowledge graphs and semantic association networks. Together, these ML technologies establish a "structure-property-process" triadic data system, providing a quantifiable and automatable alternative to traditional trial-and-error research, and pushing nanomaterials science toward a new stage of data-driven and intelligent discovery.
Nevertheless, the exploration of ML in the design, preparation, and performance prediction of nanomaterials is still in its infancy and faces numerous challenges. First, high-quality data is severely scarce and exhibits significant heterogeneity; most experimental data lacks standard formats, suffers from incompleteness, and has low reproducibility, which constrains model training effectiveness and generalization capabilities. Second, the structures of nanomaterials are highly complex; traditional feature descriptors struggle to adequately characterize heterogeneous interfaces, hollow structures, and multi-scale synergistic effects, exacerbating the limitations of the model's "black box" nature and insufficient interpretability. Furthermore, an efficient closed-loop mechanism between current AI models and experimental platforms has not yet been formed. Due to technical shortcomings in experimental equipment regarding precision control, temporal response, and high-throughput support, many AI-designed schemes are difficult to implement during the experimental verification stage. These challenges collectively constitute systemic bottlenecks across data, algorithms, and experimentation, urgently necessitating the construction of interdisciplinary, multi-platform collaborative holistic solutions.
The aforementioned challenges not only manifest as technical barriers but also severely impact model reliability and experimental operability in practical research. Extensive studies indicate that the unstructured nature and low reproducibility of experimental data cause significant accuracy degradation when training models are transferred across systems, a phenomenon particularly pronounced in frameworks such as MOFs and multi-component heterogeneous structures; the high-dimensional coupling of structural complexity and reaction conditions often renders design results from generative models chemically infeasible, making them difficult to translate into executable synthesis pathways; furthermore, deficiencies in feedback speed, linkage capabilities, and parameter control within experimental platforms hinder the deployment of AI-assisted strategies in real-world synthesis routes. These issues reflect a significant "decoupling" between machine learning and practical nanomaterial scenarios, urgently necessitating the establishment of synergistic pathways among data, algorithms, and experiments to propel their evolution from tool-level applications to deep, system-level integration.
Looking ahead, to further unleash the potential of ML in nanomaterials science, systematic breakthroughs are needed in data system construction, physics-informed modeling, and experimental platform integration. On one hand, it is essential to promote the establishment of standardized databases covering various types of nanomaterials and synthesis processes, encouraging data sharing within the field to enhance data scale, consistency, and reliability. On the other hand, developing interpretable algorithms incorporating physical constraints, hybrid modeling methods, and pre-training strategies tailored for small samples will become key research directions, particularly significant for multi-scale mechanism analysis and dynamic process modeling. Meanwhile, developing intelligent experimental platforms equipped with real-time feedback and autonomous control capabilities to translate AI outputs into executable process parameters will realize a closed-loop research workflow from computational design to experimental validation. In summary, AI is propelling nanomaterials research into a new era characterized by the deep integration of intelligence, automation, and sustainability; its continued development will depend on the profound fusion and evolution of data, algorithms, and experimental systems.

References

Publishing order | Descend order by publishing year | Descend order by cited within
[1]
Sealfon R S G, Wong A K, Troyanskaya O G. Nat. Rev. Mater., 2021, 6(8): 717.

[2]
Xiouras C, Cameli F, Quilló G L, Kavousanakis M E, Vlachos D G, Stefanidis G D. Chem. Rev., 2022, 122(15): 13006.

[3]
Shu W, Li J C, Liu J X, Zhu C W, Wang T R, Feng L, Ouyang R H, Li W X. J. Am. Chem. Soc., 2024, 146(12): 8737.

[4]
Aykol M, Herring P, Anapolsky A. Nat. Rev. Mater., 2020, 5(10): 725.

[5]
Park J, Sorourifar F, Muthyala M R, Houser A M, Tuttle M, Paulson J A, Zhang S Y. J. Am. Chem. Soc., 2024, 146(45): 31230.

[6]
Wang X Y, Wang Z Y, Gao P Y, Zhang C Q, Lv J, Wang H, Liu H F, Wang Y C, Ma Y M. Nat. Commun., 2023, 14: 2924.

[7]
Qiao Y, Hu R, Gu Y, Tang F J, Luo S H, Zhang H T, Tian J H, Cheng J, Tian Z Q. Sci. China Chem., 2024, 54(3): 338.

[8]
Park W, Shin H, Choi B, Rhim W K, Na K, Han D K. Prog. Mater. Sci., 2020, 114: 100686.

[9]
Mao J H, Shen Z H, Wang J, Liu R L, Liu X F, Lan Y, Zhou M J, Jiang J Y, Shen Y, Nan C W. ACS Nano, 2025, 19(19): 18397.

[10]
Pan X Y, Zhang Z Y, Yun Y, Zhang X, Sun Y L, Zhang Z X, Wang H D, Yang X, Tan Z Y, Yang Y Q, Xie H F, Bogdanov B, Zmaga G, Senyushkin P, Wei X M, Song Y L, Su M. J. Am. Chem. Soc., 2024, 146(28): 19239.

[11]
Yan X L, Yue T T, Winkler D A, Yin Y G, Zhu H, Jiang G B, Yan B. Chem. Rev., 2023, 123(13): 8575.

[12]
Yi Z G, Gao H X, Ji X L, Yeo X Y, Chong S Y, Mao Y J, Luo B W, Shen C, Han S Y, Wang J W, Jung S, Shi P, Ren H L, Liu X G. J. Am. Chem. Soc., 2021, 143(36): 14907.

[13]
Kim J, Rhee D, Song O, Kim M, Kwon Y H, Lim D U, Kim I S, Mazánek V, Valdman L, Sofer Z, Cho J H, Kang J. Adv. Mater., 2022, 34(12): 2106110.

[14]
Willner M R, Vikesland P J. J. Nanobiotechnol., 2018, 16: 95.

[15]
Levell Z, Le J B, Yu S, Wang R Y, Ethirajan S, Rana R, Kulkarni A, Resasco J, Lu D Y, Cheng J, Liu Y Y. Chem. Rev., 2024, 124(14): 8620.

[16]
Mashhadimoslem H, Ali Abdol M, Zanganeh K, Shafeen A, AlHammadi A A, Kamkar M, Elkamel A. ACS Appl. Energy Mater., 2024, 7(19): 8596.

[17]
Omidvar M, Zhang H F, Ihalage A A, Saunders T G, Giddens H, Forrester M, Haq S, Hao Y. Nat. Commun., 2024, 15: 6554.

[18]
Szymanski N J, Nevatia P, Bartel C J, Zeng Y, Ceder G. Nat. Commun., 2023, 14: 6956.

[19]
Liu Z P, Zhou X F, Qian Y T. Adv. Mater., 2010, 22(17): 1963.

[20]
Tao H C, Wu T Y, Aldeghi M, Wu T C, Aspuru-Guzik A, Kumacheva E. Nat. Rev. Mater., 2021, 6(8): 701.

[21]
Wan K W, Wang H, Shi X H. ACS Nano, 2024, 18(19): 12367.

[22]
Butler K T, Davies D W, Cartwright H, Isayev O, Walsh A. Nature, 2018, 559(7715): 547.

[23]
Song T, Luo M, Zhang X L, Chen L J, Huang Y, Cao J Q, Zhu Q, Liu D B, Zhang B C, Zou G, Zhang G Q, Zhang F, Shang W W, Fu Y, Jiang J, Luo Y. J. Am. Chem. Soc., 2025, 147(15): 12534.

[24]
Tom G, Schmid S P, Baird S G, Cao Y, Darvish K, Hao H, Lo S, Pablo-García S, Rajaonson E M, Skreta M, Yoshikawa N, Corapi S, Akkoc G D, Strieth-Kalthoff F, Seifrid M, Aspuru-Guzik A. Chem. Rev., 2024, 124(16): 9633.

[25]
López C. Adv. Mater., 2023, 35(23): 55.

[26]
Ares N. Nat. Rev. Mater., 2021, 6(10): 870.

[27]
Fiedler L, Modine N A, Schmerler S, Vogel D J, Popoola G A, Thompson A P, Rajamanickam S, Cangi A. NPJ Comput. Mater., 2023, 9: 115.

[28]
Jiang J Q, Chen M K, Fan J A. Nat. Rev. Mater., 2021, 6(8): 679.

[29]
Pyzer-Knapp E O, Pitera J W, Staar P W J, Takeda S, Laino T, Sanders D P, Sexton J, Smith J R, Curioni A. NPJ Comput. Mater., 2022, 8: 84.

[30]
von Lilienfeld O A, Müller K R, Tkatchenko A. Nat. Rev. Chem., 2020, 4(7): 347.

[31]
Sha W X, Guo Y Q, Yuan Q, Tang S, Zhang X F, Lu S F, Guo X, Cao Y C, Cheng S J. Adv. Intell. Syst., 2020, 2(4): 1900143.

[32]
Jia Y Y, Hou X, Wang Z W, Hu X G. ACS Sustainable Chem. Eng., 2021, 9(18): 6130.

[33]
Reis M, Gusev F, Taylor N G, Chung S H, Verber M D, Lee Y Z, Isayev O, Leibfarth F A. J. Am. Chem. Soc., 2021, 143(42): 17677.

[34]
Du M H, Dai Y H, Jiang L P, Su Y M, Qi M Q, Wang C, Long L S, Zheng L S, Kong X J. J. Am. Chem. Soc., 2023, 145(42): 23188.

[35]
Mashhadimoslem H, Ali Abdol M, Karimi P, Zanganeh K, Shafeen A, Elkamel A, Kamkar M. ACS Nano, 2024, 18(35): 23842.

[36]
Wang Y, Jiang Z J, Wang D R, Lu W G, Li D. J. Am. Chem. Soc., 2024, 146(10): 6955.

[37]
Guo H Z, Lu Y H, Lei Z D, Bao H, Zhang M W, Wang Z M, Guan C T, Tang B J, Liu Z, Wang L. Nat. Commun., 2024, 15: 4843.

[38]
Kang Y, Lee W, Bae T, Han S, Jang H, Kim J. J. Am. Chem. Soc., 2025, 147(5): 3943.

[39]
Nandy A, Duan C R, Kulik H J. J. Am. Chem. Soc., 2021, 143(42): 17535.

[40]
Xie Y C, Zhang C, Hu X Q, Zhang C, Kelley S P, Atwood J L, Lin J. J. Am. Chem. Soc., 2020, 142(3): 1475.

[41]
Zhao Y M, Zhao Y J, Wang J, Wang Z. Ind. Eng. Chem. Res., 2025, 64(9): 4637.

[42]
Zheng Z L, Alawadhi A H, Chheda S, Neumann S E, Rampal N, Liu S C, Nguyen H L, Lin Y H, Rong Z C, Siepmann J I, Gagliardi L, Anandkumar A, Borgs C, Chayes J T, Yaghi O M. J. Am. Chem. Soc., 2023, 145(51): 28284.

[43]
Abed J, Heras-Domingo J, Sanspeur R Y, Luo M C, Alnoush W, Meira D M, Wang H, Wang J, Zhou J G, Zhou D J, Fatih K, Kitchin J R, Higgins D, Ulissi Z W, Sargent E H. J. Am. Chem. Soc., 2024, 146(23): 15740.

[44]
Bonke S A, Trezza G, Bergamasco L, Song H W, Rodríguez-Jiménez S, Hammarström L, Chiavazzo E, Reisner E. J. Am. Chem. Soc., 2024, 146(22): 15648.

[45]
Cui C X, Shen Y X, He J R, Fu Y, Hong X, Wang S, Jiang J, Luo Y. J. Am. Chem. Soc., 2024, 146(50): 34551.

[46]
Duan X L, Li Y, Zhao J H, Zhang M Y, Wang X P, Zhang L, Ma X X, Qu Y, Zhang P F. J. Am. Chem. Soc., 2025, 147(1): 651.

[47]
Li W, Liang T G, Lin Y C, Wu W X, Li S. Prog. Chem., 2022, 34(12): 2619.

(李炜, 梁添贵, 林元创, 吴伟雄, 李松. 化学进展, 2022, 34(12): 2619.)

[48]
Zhang X W, Hou F, Liu R C, Wang L, Li G Z. Prog. Chem., 2024, 36(4): 471.

(张香文, 侯放, 刘睿宸, 王莅, 李国柱. 化学进展, 2024, 36(4): 471.)

[49]
Liu Z D, Pan J J, Liu Q B. Prog. Chem., 2023, 35(4): 577.

(刘振东, 潘嘉杰, 刘全兵. 化学进展, 2023, 35(4): 577.)

[50]
Wang Z H, Wu Y C, Wu Z S, Zhu R R, Yang Y, Wu F H. Prog. Chem., 2023, 35(10): 1505.

(汪忠华, 吴亦初, 吴中山, 朱冉冉, 杨阳, 吴范宏. 化学进展, 2023, 35(10): 1505.)

[51]
Masson J F, Biggins J S, Ringe E. Nat. Nanotechnol., 2023, 18(2): 111.

[52]
Nandipati M, Fatoki O, Desai S. Materials, 2024, 17(7): 1621.

[53]
Pacchioni G. Nat. Rev. Mater., 2021, 6(8): 648.

[54]
Lu Z C, Ma D, Liu X J, Lu Z P. Commun. Mater., 2024, 5: 76.

[55]
Martini A, Hursán D, Timoshenko J, Rüscher M, Haase F, Rettenmaier C, Ortega E, Etxebarria A, Roldan Cuenya B. J. Am. Chem. Soc., 2023, 145(31): 17351.

[56]
Maulana A L, Han S, Shan Y, Chen P C, Lizandara-Pueyo C, Se B, De S, Schierle-Arndt K, Yang P D. J. Am. Chem. Soc., 2025, 147(12): 10268.

[57]
Mok D H, Back S. J. Am. Chem. Soc., 2024, 146(49): 33712.

[58]
Wang X J, Ye S, Hu W, Sharman E, Liu R, Liu Y, Luo Y, Jiang J. J. Am. Chem. Soc., 2020, 142(17): 7737.

[59]
Zhao X G, Yang Q, Xu Y, Liu Q Y, Li Z Y, Liu X X, Zhao Y X, He S G. J. Am. Chem. Soc., 2024, 146(18): 12485.

[60]
Swain M C, Cole J M. J. Chem. Inf. Model., 2016, 56(10): 1894.

[61]
Castro Nascimento C M, Pimentel A S. J. Chem. Inf. Model., 2023, 63(6): 1649.

[62]
Chen Y Q, Yu T, Song Z Q, Wang C Y, Luo J T, Xiao Y, Qiu H, Wang Q Q, Jin H M. J. Chem. Inf. Model., 2025, 65(7): 3370.

[63]
Ito S, Muraoka K, Nakayama A. Chem. Mater., 2025, 37(7): 2447.

[64]
Jin C, Guo S Y, Zhou S G, Guan J H. J. Chem. Inf. Model., 2025, 65(11): 5438.

[65]
Lai Z H, Dai G L, Lu Y H, Xu Y, Li J G, Zhu X. J. Phys. Chem. Lett., 2025, 16(20): 4962.

[66]
Rojas A J. J. Chem. Educ., 2024, 101(5): 1959.

[67]
Ruff E F, Engen M A, Franz J L, Mauser J F, West J K, Zemke J M O. J. Chem. Educ., 2024, 101(6): 2483.

[68]
Supundrika Subasinghe S M, Gersib S G, Mankad N P. J. Chem. Educ., 2025, 102(4): 1563.

[69]
Zhu J J, Jiang J Y, Yang M Q, Ren Z J. Environ. Sci. Technol., 2023, 57(46): 17667.

[70]
Fang C, Wang Z, Guo R X, Ding Y X, Ma S C, Sun X Y. J. Am. Chem. Soc., 2025, 147(12): 10750.

[71]
Jia H L, Wang C H, Wang C, Clancy P. ACS Nano, 2023, 17(1): 453.

[72]
LeCun Y, Boser B, Denker J S, Henderson D, Howard R E, Hubbard W, Jackel L D. Neural Comput., 1989, 1(4): 541.

[73]
Lee B, Yoon S, Lee J W, Kim Y, Chang J, Yun J, Ro J C, Lee J S, Lee J H. ACS Nano, 2020, 14(12): 17125.

[74]
O’Leary W, Grumet M, Kaiser W, Bučko T, Rupp J L M, Egger D A. J. Am. Chem. Soc., 2024, 146(39): 26863.

[75]
Riesel E A, Mackey T, Nilforoshan H, Xu M K, Badding C K, Altman A B, Leskovec J, Freedman D E. J. Am. Chem. Soc., 2024, 146(44): 30340.

[76]
Trummer D, Searles K, Algasov A, Guda S A, Soldatov A V, Ramanantoanina H, Safonova O V, Guda A A, Copéret C. J. Am. Chem. Soc., 2021, 143(19): 7326.

[77]
You Q, Sun Y, Wang F, Cheng J, Tang F J. J. Am. Chem. Soc., 2025, 147(17): 14667.

[78]
Zahrt A F, Mo Y M, Nandiwale K Y, Shprints R, Heid E, Jensen K F. J. Am. Chem. Soc., 2022, 144(49): 22599.

[79]
Liu Q Y, Shang C, Liu Z P. J. Am. Chem. Soc., 2021, 143(29): 11109.

[80]
Li X T, Chen L, Shang C, Liu Z P. J. Am. Chem. Soc., 2021, 143(16): 6281.

[81]
Kang P L, Shang C, Liu Z P. Acc. Chem. Res., 2020, 53(10): 2119.

[82]
Shi Y F, Kang P L, Shang C, Liu Z P. J. Am. Chem. Soc., 2022, 144(29): 13401.

[83]
Wang F, Ma Z B, Cheng J. J. Am. Chem. Soc., 2024, 146(21): 14566.

[84]
Jiao Z R, Hu P F, Xu H F, Wang Q S. ACS Chem. Health Saf., 2020, 27(6): 316.

[85]
Schäfer C, Fojt J, Lindgren E, Erhart P. J. Am. Chem. Soc., 2024, 146(8): 5402.

[86]
Taylor C J, Pomberger A, Felton K C, Grainger R, Barecka M, Chamberlain T W, Bourne R A, Johnson C N, Lapkin A A. Chem. Rev., 2023, 123(6): 3089.

[87]
Zheng Z L, Zhang O F, Borgs C, Chayes J T, Yaghi O M. J. Am. Chem. Soc., 2023, 145(32): 18048.

[88]
Zhou Y C, Wang Y, Peijnenburg W, Vijver M G, Balraadjsing S, Dong Z M, Zhao X L, Leung K M Y, Mortensen H M, Wang Z Y, Lynch I, Afantitis A, Mu Y S, Wu F C, Fan W H. Environ. Sci. Technol., 2024, 58(34): 3328.

[89]
Nielsen M K, Ahneman D T, Riera O, Doyle A G. J. Am. Chem. Soc., 2018, 140(15): 5004.

[90]
Johnson H M, Gusev F, Dull J T, Seo Y, Priestley R D, Isayev O, Rand B P. J. Am. Chem. Soc., 2024, 146(31): 21583.

[91]
Żurański A M, Gandhi S S, Doyle A G. J. Am. Chem. Soc., 2023, 145(14): 7898.

[92]
Chen J R T, Tan E X, Tang J X, Leong S X, Hue S K X, Pun C S, Phang I Y, Ling X Y. J. Am. Chem. Soc., 2025, 147(8): 6654.

[93]
Choi S E, Jang M, Yoon S, Yoo S, Ahn J, Kim M, Kim H G, Jung Y, Park S, Kim Y S, Kim T. J. Chem. Inf. Model., 2025, 65(6): 2748.

[94]
Shen C, Zhan W K, Xin K Y, Li M Y, Sun Z Y, Cong H, Xu C, Tang J, Wu Z F, Xu B, Wei Z M, Xue C L, Zhao C, Wang Z G. Nat. Commun., 2024, 15: 2724.

[95]
Rosales S, Medina O E, Garzon N, Zapata K, Taborda E A, Ordóñez J C, Cortés F B, Franco C A. Renew. Sustain. Energy Rev., 2025, 219: 115854.

[96]
Han Y, Tang B J, Wang L, Bao H, Lu Y H, Guan C T, Zhang L, Le M Y, Liu Z, Wu M H. ACS Nano, 2020, 14(11): 14761.

[97]
Lei Z, Wang Q M. Coord. Chem. Rev., 2019, 378: 382.

[98]
Panapitiya G, Avendaño-Franco G, Ren P J, Wen X D, Li Y W, Lewis J P. J. Am. Chem. Soc., 2018, 140(50): 17508.

[99]
Mekki-Berrada F, Ren Z K, Huang T, Wong W K, Zheng F, Xie J X, Tian I P S, Jayavelu S, Mahfoud Z, Bash D, Hippalgaonkar K, Khan S, Buonassisi T, Li Q X, Wang X N. NPJ Comput. Mater., 2021, 7: 55.

[100]
Dasbach M, Pyschik M, Lehmann V, Parey K, Rhinow D, Reinhardt H M, Hampp N A. ACS Nano, 2020, 14(7): 8181.

[101]
Sun S D, Maruyama S, Li Y. J. Am. Chem. Soc., 2025, 147(8): 7103.

[102]
Xu M Z, Tang B J, Lu Y H, Zhu C, Lu Q B, Zhu C, Zheng L, Zhang J Y, Han N N, Fang W D, Guo Y X, Di J, Song P, He Y M, Kang L X, Zhang Z Y, Zhao W, Guan C T, Wang X W, Liu Z. J. Am. Chem. Soc., 2021, 143(43): 18103.

[103]
Shi Z Y, Ge Y Y, Yun Q B, Zhang H. Acc. Chem. Res., 2022, 55(24): 3581.

[104]
Liu S, Pan X T, Liu H Y. Angew. Chem. Int. Ed., 2020, 59(15): 5890.

[105]
Liu X H, Ma T T, Pinna N, Zhang J. Adv. Funct. Mater., 2017, 27(37): 1702168.

[106]
Dahl J C, Niblett S, Cho Y, Wang X Z, Zhang Y, Chan E M, Alivisatos A P. J. Am. Chem. Soc., 2023, 145(42): 23076.

[107]
Gu Y F, Wu Y N, Li L C, Chen W, Li F T, Kitagawa S. Angew. Chem. Int. Ed., 2017, 56(49): 15658.

[108]
Mohanty B, Kumari S, Yadav P, Kanoo P, Chakraborty A. Coord. Chem. Rev., 2024, 519: 216102.

[109]
Masoomi M Y, Morsali A, Dhakshinamoorthy A, Garcia H. Angew. Chem. Int. Ed., 2019, 58(43): 15188.

[110]
Goeminne R, Van Speybroeck V. J. Am. Chem. Soc., 2025, 147: 3615.

[111]
He D L, Jiang Y B, Guillén-Soler M, Geary Z, Vizcaíno-Anaya L, Salley D, Del Carmen Gimenez-Lopez M, Long D L, Cronin L. J. Am. Chem. Soc., 2024, 146(42): 28952.

Options
Outlines

/

〈 〉